This paper describes the application of techniques based on dynamic neural networks for fault diagnosis. Two architectures of dynamic neural networks are used. The better net is integrated in a state observer bank, where each net describes one system behavior. Training in closed loop is used. The method of fault detection and diagnosis is based on the definition of minimum errors (residues). These residues are calculated by comparing the plant outputs and each dynamic neural network output from the state observer bank, with and without faults. Finally, this technique is applied to a tanks system, and can be demonstrated that it is possible to reduce the observers bank, i.e., it is not necessary to use a neural network for each behavior of the system, with and without faults.